agent_loop: pre-existing loop behaviours flagged during the #318 refactor review #322

During review of #318, a refactor that changed no behaviour, Tiny Sweeper flagged the behaviours below. Each one already existed before the refactor, so they were left out of that PR. Each needs a decision and possibly a fix:

• tool_search bridge: it checks only the registry, so it does not suppress a duplicate tool_search name.
• Graceful mixed turn: it ignores ContinueLoop from a control point, and it resets no recovery counters before continue_from_queue_at_finish can start another turn.
• Dropped-tool-call nudge: it does not check the model-call budget before scheduling the nudge.
• Length stops: unterminated text-dialect blocks are dropped under a length stop, with no recovery attempt.
• Mixed structured-and-tools turns: they skip output validation.
• Promoted tools: their schemas are never prepared (prepare) the way direct tool schemas are.

Context: tinyagents harness uplift (pi/OpenClaw comparison).